Hidden Markov Model strategy doubles Sharpe ratio and cuts drawdown by 65%

·1 views

A developer built a real-time market risk scoring system using a Gaussian Hidden Markov Model (GaussianHMM) trained on five years of data from five assets, including SPY, gold, and bonds. The model identifies three hidden market regimes — Bull, Neutral, and Bear — without manual labeling, using the Baum-Welch algorithm to learn transition patterns from returns and volatility. Based on the detected regime, the strategy dynamically adjusts portfolio exposure, going fully invested in bull markets and defensive during bear phases. Backtesting results showed a Sharpe ratio of 2.12 versus 1.11 for a buy-and-hold approach, while maximum drawdown shrank from -21.9% to -7.5%. The author argues that HMMs offer an honest way to capture regime shifts that traders sense intuitively, without overfitting to exact price predictions.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer builds reproducible Erlang learning environment using Nix flakes

A developer learning Erlang has published a GitHub repository combining a Nix-based development shell with small experimental programs written during the learning process. The setup uses Nix flakes to pin the environment to a specific Nixpkgs revision, ensuring that Erlang/OTP 29.0.6 is used consistently by anyone who clones and enters the shell. The repository is structured around two components: a flake.nix configuration file and a directory of programs written while working through the book 'Learn You Some Erlang for Great Good!'. The environment is deliberately minimal, currently including only Erlang itself, with tools like Rebar3 left out until more complex projects require them. The developer notes the setup may expand over time as the scope of their Erlang experiments grows.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Most Customer Churn Models Fail — and What Actually Fixes Them

Most churn prediction programmes share a core flaw: they rank customers by likelihood of leaving rather than by how likely they are to respond positively to intervention. This means retention campaigns often waste resources on 'lost causes' who have already decided to leave, 'sure things' who would have stayed anyway, and 'sleeping dogs' who are actually pushed to cancel by being contacted. Experts recommend uplift modelling, which estimates the change in churn probability caused by an intervention, requiring randomised treatment data to measure true responsiveness. Involuntary churn from failed payments should also be separated out before model training, as it corrupts results and is better addressed through payment-retry logic and pre-expiry reminders. Additionally, data leakage from outcome-encoded features is a common reason churn models show inflated validation scores that do not hold up in production.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Per-seat AI coding tools waste idle capacity while heavy users hit limits early

Companies paying for per-seat AI coding subscriptions routinely lose unused capacity each week because allowances expire rather than roll over, while some engineers exhaust their allocation by mid-week. Usage is highly uneven across roles — engineers deep in migrations may max out for weeks, while those in meetings or non-coding roles barely touch their seats. The practical fix, according to the analysis, is redistributing work items to colleagues with spare capacity rather than sharing credentials or buying additional seats, both of which create new problems. This approach requires tasks to be clearly documented so they can be handed off, which the author notes is a prerequisite for good AI-assisted workflows in general. Any gains from this method should be counted as throughput improvements rather than cost savings, since money is only truly saved if the alternative would have been purchasing more seats.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Data Validation Catches Silent Web Scraping Failures Before They Spread

Web scrapers can return a successful HTTP status while delivering corrupted or mismatched data, making failures invisible until downstream systems are affected. A developer guide published on DEV Community recommends defining a strict data contract — including field types, required values, and accepted ranges — before scraping at scale. Using Python's Pydantic library, developers can validate each scraped record at ingestion and automatically route invalid rows to a re-scrape queue with a logged reason, rather than storing nulls silently. The guide advises running checks at three stages: before scraping to clean URL inputs, during scraping to validate extracted fields, and after scraping to monitor aggregate anomalies like sudden spikes in missing prices. Such aggregate signals often indicate a site template change, a proxy serving a bot-check page, or an unexpected regional layout rather than isolated one-off errors.

0 comments Read more at DEV Community